A low-light image enhancement method, system and readable storage medium

Abstract

The embodiments of the present application provide a low-light image enhancement method, system and readable storage medium, the method comprising acquiring a low-light image; determining a target depth curve estimation network, and performing unsupervised light enhancement processing on the low-light image through the target depth curve estimation network to obtain a corresponding brightened image; determining a target decomposition network, and performing reflectance decomposition processing on the low-light image and the brightened image respectively through the target decomposition network to obtain a target low-light reflectivity and a target enhanced image reflectivity whose corresponding reflectivity values ​​remain consistent; when it is determined that there are blurred points and / or noise points in the target reflectivity image obtained after light enhancement, the target reflectivity image is repaired through a pre-constructed target repair network to obtain the optimal image enhancement effect.

Description

Technical Field

[0001] The present application relates to the technical field of low-light image processing, and in particular to a low-light image enhancement method, system and readable storage medium. Background Art

[0002] In the past few decades, many researchers have devoted themselves to low-light image enhancement. The earliest method was to directly amplify the global illumination. Later, the dynamic range of the image was extended by the histogram equalization (HE) method, but their results may present undesirable illumination with amplified strong noise.

[0003] At present, with the widespread use of CNN (convolutional neural network) in the field of image, the first Retinex-Net model based on Retinex theory combined with CNN was born in 2018. However, in the current low-light enhancement model based on Retinex theory, how to eliminate the degradation phenomena such as noise and color distortion hidden in the dark before after brightening the dark area, and how to train a model without clearly defined ground truth lighting conditions to enhance low-light images by only looking at two / several different examples, are still not well handled. In addition, due to the shortcomings of the decomposition network, the reflectivity obtained always inevitably has the phenomenon of noise point degradation, and there is a problem of poor image enhancement effect. Summary of the invention

[0004] The purpose of the embodiments of the present application is to provide a low-light image enhancement method, system and readable storage medium, which can improve the image enhancement effect.

[0005] The present invention provides a method for enhancing low-light images, comprising the following steps:

[0006] Acquire low-light images;

[0007] Determining a target depth curve estimation network, and performing unsupervised illumination enhancement processing on the low-light image through the target depth curve estimation network to obtain a corresponding target brightened image;

[0008] Determine a target decomposition network, and perform reflectance decomposition processing on the low-light image and the brightened image respectively through the target decomposition network to obtain a target low-light reflectance and a target enhanced image reflectance whose corresponding reflectance values ​​are consistent;

[0009] When it is determined that there are blur points and / or noise points in the target reflectivity image obtained after illumination enhancement, the target reflectivity image is repaired by a pre-constructed target repair network to obtain an optimal image enhancement effect.

[0010] In a second aspect, the embodiment of the present application further provides a low-light image enhancement system, the system comprising a data acquisition module, a light enhancement module, a reflectivity decomposition processing module and an enhancement optimization module, wherein

[0011] The data acquisition module is used to acquire low-light images;

[0012] The illumination enhancement module is used to determine a target depth curve estimation network, and perform unsupervised illumination enhancement processing on the low-light image through the target depth curve estimation network to obtain a corresponding brightened image;

[0013] The reflectivity decomposition processing module is used to determine a target decomposition network, and perform reflectivity decomposition processing on the low-light image and the brightened image respectively through the target decomposition network to obtain a target low-light reflectivity and a target enhanced image reflectivity whose corresponding reflectivity values ​​remain consistent;

[0014] The enhancement optimization module is used to, when it is determined that there are blur points and / or noise points in the target reflectivity image obtained after illumination enhancement, perform restoration processing on the target reflectivity image through a pre-constructed target restoration network to obtain the optimal image enhancement effect.

[0015] In a third aspect, an embodiment of the present application further provides a readable storage medium, wherein the readable storage medium includes a low-light image enhancement method program. When the low-light image enhancement method program is executed by a processor, the steps of a low-light image enhancement method as described in any one of the above items are implemented.

[0016] As can be seen from the above, the low-light image enhancement method, system and readable storage medium provided by the embodiments of the present application utilize the powerful deep learning capabilities of the decomposition network and the repair network to jointly train the reflectance obtained by decomposition under different illuminations. On the one hand, the similarity constraint function is constructed by utilizing the characteristic that the reflectance of pictures with different illuminations should remain consistent. While ensuring the accuracy of illumination enhancement, the reflectance of the cleanest bottom layer of the picture is learned through the fitting ability of deep learning. On the other hand, when it is determined that there are blurred points and / or noise points in the target reflectance image obtained after illumination enhancement, the target reflectance image obtained after illumination enhancement is repaired and optimized through the repair network, so as to obtain a higher quality reflectance and improve the image enhancement effect.

[0017] Other features and advantages of the present application will be described in the following description, and partly become apparent from the description, or be understood by practicing the embodiments of the present application. The purpose and other advantages of the present application can be realized and obtained by the structures specifically pointed out in the written description, claims, and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments of the present application will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0019] Figure 1 A flowchart of a low-light image enhancement method provided in an embodiment of the present application;

[0020] Figure 2 A schematic diagram of the structure of a depth curve estimation network provided in an embodiment of the present application;

[0021] Figure 3 A schematic diagram of the structure of a decomposed network provided in an embodiment of the present application;

[0022] Figure 4 A schematic diagram of the structure of a repair network provided in an embodiment of the present application;

[0023] Figure 5 A schematic diagram of the structure of the overall network system provided in the embodiment of the present application;

[0024] Figure 6 A schematic diagram of the structure of a low-light image enhancement system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0025] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. The components of the embodiments of the present application described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the application claimed for protection, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present application.

[0026] It should be noted that similar reference numerals and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. At the same time, in the description of this application, the terms "first", "second", etc. are only used to distinguish the description and cannot be understood as indicating or implying relative importance.

[0027] Please refer to Figure 1 , Figure 1This is a flowchart of a low-light image enhancement method in some embodiments of the present application. Taking the method applied to a computer device (the computer device may be a terminal or a server, and the terminal may be, but is not limited to, various personal computers, laptops, smart phones, tablet computers, and portable wearable devices. The server may be an independent server or a server cluster composed of multiple servers) as an example, the method includes the following steps:

[0028] Step S100: acquiring a low-light image.

[0029] Step S200: determine a target depth curve estimation network, and perform unsupervised illumination enhancement processing on the low-light image through the target depth curve estimation network to obtain a corresponding brightened image.

[0030] Step S300, determining a target decomposition network, and performing reflectance decomposition processing on the low-light image and the brightened image respectively through the target decomposition network to obtain a target low-light reflectance and a target enhanced image reflectance whose corresponding reflectance values ​​remain consistent.

[0031] Step S400, when it is determined that there are blur points and / or noise points in the target reflectivity image obtained after illumination enhancement, the target reflectivity image is repaired by a pre-constructed target repair network to obtain an optimal image enhancement effect.

[0032] As can be seen from the above, a low-light image enhancement method provided by an embodiment of the present application utilizes the powerful deep learning capabilities of the decomposition network and the repair network to jointly train the reflectance obtained by decomposition under different illuminations. On the one hand, the reflectance of pictures with different illuminations should remain consistent to construct a similarity constraint function. While ensuring the accuracy of illumination enhancement, the reflectance of the cleanest bottom layer of the picture is learned through the fitting ability of deep learning. On the other hand, when it is determined that there are blurred points and / or noise points in the target reflectance image obtained after illumination enhancement, the target reflectance image obtained after illumination enhancement is repaired and optimized through the repair network, so as to obtain a higher quality reflectance and improve the image enhancement effect.

[0033] In one embodiment, in step S200, unsupervised illumination enhancement processing is performed on the low-light image through a target depth curve estimation network to obtain a corresponding brightened image, including:

[0034] Step S2001: taking the low-light image as the input of the target depth curve estimation network, and processing the input image through the target depth curve estimation network to obtain the corresponding optimal fitting light enhancement curve.

[0035] For details, please refer to Figure 2When the computer device performs illumination enhancement processing on the input image, it uses a deep curve estimation network including a five-layer convolutional network to obtain a set of best fitting light enhancement curves LE corresponding to the input image.

[0036] Step S2002, mapping all pixels included in the color channel of the input image by iteratively applying the best fitting light enhancement curve to obtain a corresponding brightened image.

[0037] Specifically, the above-mentioned color channels can be RGB channels (i.e., red, green, and blue channels). It should be noted that the RGB channel is an additive color model, which adds the three primary colors of red (Red), green (Green), and blue (Blue) in different proportions to synthesize various color lights.

[0038] In one embodiment, the computer device maps all pixels in the input image by iteratively applying the obtained optimal fitting light enhancement curve to obtain a final brightened image. In order to ensure the quality of the enhancement map, during the network model training process, multiple loss functions for the network model can also be used to perform loss constraints to ensure the accuracy of the light enhancement.

[0039] In the above embodiment, the acquired low-light image is preprocessed with unsupervised illumination enhancement through a deep curve estimation network to obtain a brightened image after illumination enhancement, wherein the brightened image can be used as a standard illumination image paired during supervised learning and can be used to supervise subsequent decomposition effects.

[0040] In one embodiment, in step S2001, determining a target depth curve estimation network includes:

[0041] Step S20011, constructing an initial depth curve estimation network, wherein the initial depth curve estimation network is composed of a multi-layer convolutional network.

[0042] Step S20012, obtaining a low-light training image for network training, and performing illumination enhancement training processing on the low-light training image through an initial depth curve estimation network, wherein during the training process, the initial depth curve estimation network is constrained according to a pre-constructed first target loss function to ensure the accuracy of illumination enhancement, and the first target loss function includes a spatial consistency loss function, an exposure control loss function, a color consistency loss function, and an illumination smoothness loss function.

[0043] Step S20013, when the preset first training end condition is reached, output and determine the required target depth curve estimation network.

[0044] In the above embodiment, a loss function is constructed to impose loss constraints on the depth curve estimation network, thereby ensuring the accuracy of illumination enhancement, providing a good data basis for subsequent reflectivity decomposition, and ensuring the accuracy of data analysis.

[0045] In one embodiment, the first objective loss function includes:

[0046] The spatial consistency loss function is determined based on the local differences between the brightened image obtained during training and the low-light training image.

[0047] The exposure control loss function is determined based on the average intensity value generated by the local area in the brightened image obtained during training.

[0048] The color consistency loss function is determined based on the average intensity value generated by the corresponding color channel in the brightened image obtained during training.

[0049] The lighting smoothness loss function is determined based on the results obtained by integrating the horizontal gradient operation and the vertical gradient operation.

[0050] In one embodiment, the reference formula of the spatial consistency loss function includes:

[0051]

[0052] Among them, L spa is the spatial consistency loss function, K is the total number of local regions, Ω(i) is the four adjacent regions centered on i, and Y and I represent the brightened image obtained through training and the low-light training image, respectively.

[0053] The reference formula of the exposure control loss function includes

[0054]

[0055] Among them, L exp is the exposure control loss function, M is the total number of non-overlapping local regions, and Y k It is the intensity value generated by the local area k in the brightened image obtained during training.

[0056] The reference formula of the color consistency loss function includes:

[0057]

[0058] Among them, L col is the color consistency loss function, J p , J q They are the intensity values ​​generated by the corresponding channels p and q in the brightened images obtained during training.

[0059] The reference formula of the lighting smoothness loss function includes:

[0060]

[0061] Among them, L tvA is the lighting smoothness loss function, A is the reference image of the fitted light enhancement curve, N is the preset maximum number of iterations, and They are horizontal gradient operation and vertical gradient operation respectively.

[0062] In one embodiment, in step S300, determining a target decomposition network includes:

[0063] Step S3001, construct an initial decomposition network, which includes a reflectance branch network corresponding to reflection processing and an illumination branch network corresponding to illumination processing, wherein the reflectance branch network is composed of a multi-layer U-Net network including convolutional layers and sigmoid layers, and the illumination branch network is composed of two conv+Relu layers and one conv layer on the cascade feature map of the reflectance branch network.

[0064] Specifically, the structural diagram of the decomposition network can be referred to Figure 3 .

[0065] In one of the embodiments, according to the Retinex theory, each image is composed of a reflectance R and an illumination map I. Therefore, when designing the network structure, it is considered to decompose the decomposition network into a reflection branch corresponding to the reflectance processing and an illumination branch corresponding to the illumination processing. In one of the embodiments, the reflectance branch adopts a typical five-layer U-Net network, wherein the U-Net network includes a convolution layer and a sigmoid layer. The illumination branch is composed of two conv+Relu layers and one conv layer on the cascade feature map of the reflectance branch, and finally the processing result is output by the sigmoid layer.

[0066] Step S3002, obtaining a low-light training image and a brightened training image output from the training after loss, and performing reflectance decomposition training processing on the low-light training image and the brightened training image through an initial decomposition network. During the training process, a similarity constraint loss function is constructed based on the principle that the reflectance between images of different illuminations remains consistent, and the parameters of the network model are adjusted to ensure the consistency of the values ​​between the obtained low-light training reflectance and the corresponding enhanced training reflectance.

[0067] Specifically, since the reflectivity of pictures with different illumination levels should remain consistent, in the current embodiment, it is used as a constraint condition, and a similarity constraint loss function is constructed based on the learned reflectivity recognition results of pictures with different illumination levels. The model parameters are optimized through continuous gradient iterations to learn the reflectivity of the cleanest bottom layer of the picture through the fitting ability of deep learning.

[0068] Step S3003: when the preset second training end condition is reached, output and determine the required target decomposition network.

[0069] The above embodiment, combined with the fitting ability of deep learning, learns the reflectivity of the cleanest bottom layer of the image, which can effectively improve the recognition accuracy of the reflectivity of the enhanced image, ensure the analysis accuracy, and improve the image enhancement effect.

[0070] In one embodiment, in step S3002, during the process of training the initial decomposition network, the method further includes:

[0071] Step S30021, construct a second objective loss function to constrain the initial decomposition network through the second objective loss function to avoid reducing the value accuracy under the influence of lighting, wherein the second objective loss function includes a reconstruction loss function, a local stage smoothness loss function, and a reflectivity gradient contrast loss function.

[0072] Specific:

[0073] (1) The calculation formula for reconstruction loss degree can refer to:

[0074] L recon =||SR·I||;

[0075] Among them, L recon is the reconstruction loss function, S is the decomposed image, R is the decomposed reflectivity, and I is the decomposed illumination map.

[0076] (2) The calculation formula for the degree of local stage smoothness loss can be referred to:

[0077]

[0078] Among them, L R is the local stage smooth loss function, c represents the R, G, and B channels respectively, F(*) represents the histogram equalization operator of the image “*”, and λ is the preset weight parameter.

[0079] (3) The calculation formula of the gradient contrast loss of reflectivity can be referred to as:

[0080]

[0081] In summary, the total loss function of the decomposition network disclosed in the embodiment of the present application can be expressed as:

[0082] L decom =L recon +λ1L R +λ2L I ;

[0083] Among them, on the one hand, the embodiment of the present application does not limit the specific value of the weight parameter. In different embodiments, it can be adjusted in combination with the network training situation, but its limited range is between 0 and 1. On the other hand, the above-mentioned total loss function means that the maximum channel of the reflectivity should conform to the maximum channel of the low-light image and have the maximum entropy.

[0084] In one of the embodiments, the target restoration network is composed of nine convolutional units, wherein each convolutional unit includes multiple conv+Relu layers and a max pooling layer; during the training of the pre-constructed initial restoration network, the method further includes: performing network constraints by constructing a third target loss function based on the degree of difference between the reflectivity of the obtained network restoration and the reflectivity of the brightened image to ensure the accuracy of image restoration, and outputting and determining the required target restoration network when the preset training end condition is reached.

[0085] For details, please refer to Figure 4 and Figure 5 The principle of image restoration can be understood as follows: the computer device uses the enhanced reflectance map to constrain the low light to obtain the restoration of the reflectance map, and at the same time inputs the corresponding illumination map to assist its restoration to obtain a higher quality reflectance map.

[0086] In one embodiment, the reference calculation formula of the third objective loss function includes:

[0087]

[0088] in, is the reflectivity of the network repair, R h is the reflectance of the brightened image for comparison, is the gradient operator, and SSIM(*) is used to calculate the structural similarity of “*”.

[0089] In the above embodiment, a joint training model is established for the reflectance decomposed under different illuminations, and the decomposition of pictures with different illuminations is constrained by structural similarity. Then, the obtained reflectance is repaired and optimized. Driven by data from a large number of data sets, a deep learning network can be trained that combines the consistency of reflectance to jointly approximate the optimal reflectance. Compared with the model that learns the optimal reflectance with a single reflectance, a higher quality enhancement effect can be achieved.

[0090] Please refer to Figure 6 As shown, the embodiment of the present application also provides a low-light image enhancement system 600, which includes a data acquisition module 601, an illumination enhancement module 602, a reflectivity decomposition processing module 603 and an enhancement optimization module 604, wherein

[0091] The data acquisition module 601 is used to acquire low-light images.

[0092] The illumination enhancement module 602 is used to determine a target depth curve estimation network, and perform unsupervised illumination enhancement processing on the low-light image through the target depth curve estimation network to obtain a corresponding brightened image.

[0093] The reflectivity decomposition processing module 603 is used to determine the target decomposition network, and perform reflectivity decomposition processing on the low-light image and the brightened image through the target decomposition network, so as to obtain the target low-light reflectivity and the target enhanced image reflectivity with the corresponding reflectivity values ​​kept consistent.

[0094] The enhancement optimization module 604 is used to, when it is determined that there are blur points and / or noise points in the target reflectivity image obtained after illumination enhancement, perform restoration processing on the target reflectivity image through a pre-constructed target restoration network to obtain the optimal image enhancement effect.

[0095] The above-mentioned low-light image enhancement system uses the powerful deep learning capabilities of the decomposition network and the restoration network to jointly train the reflectance obtained by decomposition under different illuminations. On the one hand, the similarity constraint function is constructed by using the characteristic that the reflectance of pictures with different illuminations should remain consistent. While ensuring the accuracy of illumination enhancement, the reflectance of the cleanest bottom layer of the picture is learned through the fitting ability of deep learning. On the other hand, when it is determined that there are blurred points and / or noise points in the target reflectance image obtained after illumination enhancement, the target reflectance image obtained after illumination enhancement is restored and optimized through the restoration network, so as to obtain a higher quality reflectance and improve the image enhancement effect.

[0096] The embodiment of the present application provides a readable storage medium, and when the computer program is executed by the processor, the method in any optional implementation of the above embodiment is executed. Wherein, the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (Static Random Access Memory, referred to as SRAM), electrically erasable programmable read-only memory (Electrically Erasable Programmable Read-Only Memory, referred to as EEPROM), erasable programmable read-only memory (Erasable Programmable Read Only Memory, referred to as EPROM), programmable read-only memory (Programmable Red-Only Memory, referred to as PROM), read-only memory (Read-Only Memory, referred to as ROM), magnetic memory, flash memory, disk or optical disk.

[0097] The above-mentioned readable storage medium uses the powerful deep learning capabilities of the decomposition network and the repair network to jointly train the reflectance obtained by decomposition under different illuminations. On the one hand, the reflectance of pictures with different illuminations should remain consistent to construct a similarity constraint function. While ensuring the accuracy of illumination enhancement, the reflectance of the cleanest bottom layer of the picture is learned through the fitting ability of deep learning. On the other hand, when it is determined that there are blurred points and / or noise points in the target reflectance image obtained after illumination enhancement, the target reflectance image obtained after illumination enhancement is repaired and optimized through the repair network, so as to obtain a higher quality reflectance and improve the image enhancement effect.

[0098] In the embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.

[0099] In addition, the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0100] Furthermore, the functional modules in the various embodiments of the present application may be integrated together to form an independent part, or each module may exist separately, or two or more modules may be integrated to form an independent part.

[0101] In this document, relational terms such as first and second, etc. are used merely to distinguish one entity or operation from another entity or operation, but do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0102] The above description is only an embodiment of the present application and is not intended to limit the protection scope of the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A low-light image enhancement method, characterized in that: The following steps are involved: Acquire low-light images; Determine a target depth curve estimation network, and perform unsupervised illumination enhancement processing on the low-light image through the target depth curve estimation network to obtain a corresponding brightened image; Determine a target decomposition network, and perform reflectance decomposition processing on the low-light image and the brightened image respectively through the target decomposition network to obtain a target low-light reflectance and a target enhanced image reflectance whose corresponding reflectance values ​​are consistent; When it is determined that there are blur points and / or noise points in the target reflectivity image obtained after illumination enhancement, the target reflectivity image is repaired by a pre-constructed target repair network to obtain an optimal image enhancement effect; The determining of the target decomposition network comprises: Constructing an initial decomposition network, the initial decomposition network comprising a reflectance branch network corresponding to reflection processing and an illumination branch network corresponding to illumination processing, wherein the reflectance branch network is composed of a multi-layer U-Net network composed of a convolutional layer and a sigmoid layer, and the illumination branch network is composed of two conv+Relu layers and one conv layer on a cascade feature map of the reflectance branch network; Obtain a low-light training image and a brightened training image output from the training after loss, and perform reflectance decomposition training processing on the low-light training image and the brightened training image through the initial decomposition network, wherein during the training process, a similarity constraint loss function is constructed in combination with the principle that the reflectance between images of different illuminations remains consistent, and the parameters of the network model are adjusted to ensure the consistency of the values ​​of the obtained low-light training reflectance and the corresponding enhanced training reflectance; When the preset second training end condition is reached, output and determine the required target decomposition network; The target repair network is composed of a plurality of convolutional units, wherein each convolutional unit includes a plurality of conv+Relu layers and a max pooling layer; in the process of training the pre-constructed initial repair network, the method further includes: The network is constrained by a third objective loss function constructed according to the degree of difference between the reflectivity of the network restoration and the reflectivity of the brightened image to ensure the accuracy of image restoration. When the preset training end condition is reached, the required target restoration network is output and determined.

2. The method according to claim 1, characterized in that The step of performing unsupervised illumination enhancement processing on the low-light image through the target depth curve estimation network to obtain a corresponding brightened image includes: Using the low-light image as input to the target depth curve estimation network, and processing the input image through the target depth curve estimation network to obtain a corresponding optimal fitting light enhancement curve; All pixels contained in the color channel of the input image are mapped by iteratively applying the best fitting light enhancement curve to obtain a corresponding brightened image.

3. The method according to claim 2, characterized in that The target depth curve estimation network is determined, comprising: Constructing an initial depth curve estimation network, wherein the initial depth curve estimation network includes a multi-layer convolutional network; Acquire a low-light training image for network training, and perform illumination enhancement training processing on the low-light training image through the initial depth curve estimation network, wherein during the training process, the initial depth curve estimation network is constrained according to a pre-constructed first target loss function to ensure the accuracy of illumination enhancement, and the first target loss function includes a spatial consistency loss function, an exposure control loss function, a color consistency loss function, and an illumination smoothness loss function; When the preset first training end condition is reached, the required target depth curve estimation network is output and determined.

4. The method according to claim 3, characterized in that The first objective loss function includes: A spatial consistency loss function determined by the local differences between the brightened image obtained during training and the low-light training image; An exposure control loss function determined based on the average intensity value generated in the local area of ​​the brightened image obtained during training; A color consistency loss function determined by the average intensity value of the corresponding color channel in the brightened image obtained during training; The lighting smoothness loss function is determined based on the results obtained by integrating the horizontal gradient operation and the vertical gradient operation.

5. The method according to claim 4, characterized in that The reference formula of the spatial consistency loss function includes: Among them, L spa is the spatial consistency loss function, K is the total number of local regions, Ω(i) is the four adjacent regions centered on i, and Y and I represent the brightened image obtained through training and the low-light training image, respectively; The reference formula of the exposure control loss function includes Among them, L exp is the exposure control loss function, M is the total number of non-overlapping local regions, and Y k is the intensity value generated by the local area k in the brightened image obtained through training; The reference formula of the color consistency loss function includes: Among them, L col is the color consistency loss function, J p , J q are the intensity values ​​generated by the corresponding channels p and q in the brightened images obtained during training; The reference formula of the lighting smoothness loss function includes: Among them, L tvA is the lighting smoothness loss function, A is the reference image of the fitted light enhancement curve, N is the preset maximum number of iterations, and They are horizontal gradient operation and vertical gradient operation respectively.

6. The method according to claim 1, characterized in that During the training of the initial decomposition network, the method further comprises: Construct a second objective loss function to constrain the initial decomposition network through the second objective loss function to avoid reducing the value accuracy under the influence of lighting, wherein the second objective loss function includes a reconstruction loss function, a local stage smoothness loss function, and a reflectivity gradient contrast loss function.

7. A low-light image enhancement system, applied to the low-light image enhancement method according to claim 1, characterized in that: The system includes a data acquisition module, an illumination enhancement module, a reflectivity decomposition processing module and an enhancement optimization module, wherein The data acquisition module is used to acquire low-light images; The illumination enhancement module is used to determine a target depth curve estimation network, and perform unsupervised illumination enhancement processing on the low-light image through the target depth curve estimation network to obtain a corresponding brightened image; The reflectivity decomposition processing module is used to determine a target decomposition network, and perform reflectivity decomposition processing on the low-light image and the brightened image respectively through the target decomposition network to obtain a target low-light reflectivity and a target enhanced image reflectivity whose corresponding reflectivity values ​​remain consistent; The enhancement optimization module is used to, when it is determined that there are blur points and / or noise points in the target reflectivity image obtained after illumination enhancement, perform restoration processing on the target reflectivity image through a pre-constructed target restoration network to obtain the optimal image enhancement effect.

8. A readable storage medium, characterized in that: The readable storage medium includes a low-light image enhancement method program, and when the low-light image enhancement method program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images